Synthetic I/Q

Unlike over-the-air captures that require manual annotation, synthetic I/Q is generated directly from a known modulation scheme and bit sequence. This eliminates human labeling error, providing a 100% accurate ground truth for supervised learning. Every sample is paired with its exact modulation type, symbol rate, and signal-to-noise ratio (SNR), enabling precise loss calculation during neural network training.

Synthetic generation bypasses the logistical and legal constraints of real-world spectrum collection. Engineers can programmatically create unlimited volumes of training data, covering rare waveforms like MIL-STD-188-110 serial tone modems or custom proprietary protocols. This scalability is critical for training deep neural networks that require millions of diverse examples to generalize effectively without overfitting.

Synthetic I/Q allows for isolated, deterministic control over every channel impairment. This enables curriculum learning, where models are first trained on clean AWGN and progressively exposed to complex fading models.

• Carrier Frequency Offset (CFO): Precise phase rotation applied mathematically.
• Multipath Fading: Rayleigh or Rician profiles with configurable delay spreads.
• Non-Linear Distortion: Simulated power amplifier (PA) non-linearity via Saleh or Rapp models.

Real-world spectrum is dominated by a few common waveforms (e.g., QPSK, 16QAM), leading to severe class imbalance in collected datasets. Synthetic generation ensures a perfectly uniform prior distribution across all target classes. This prevents the classifier from developing a bias toward high-density signal types and ensures robust performance on rare, high-value signals like 256QAM or spread-spectrum waveforms.

Synthetic I/Q is not limited to software simulation. The digital samples can be streamed through an arbitrary waveform generator (AWG) and upconverted to RF, creating a physical, repeatable test signal. This bridges the gap between pure simulation and real-world hardware testing, allowing engineers to validate the entire receiver pipeline—from antenna to classifier output—with a known, controllable stimulus.

By randomizing nuisance parameters during generation—such as pulse shaping filter roll-off, symbol timing offset, and burst length—synthetic data forces the neural network to learn invariant features of the modulation itself. This domain randomization technique prevents the model from latching onto spurious correlations in the training data and significantly improves generalization when deployed on real, unseen hardware.

The process of applying mathematical models of fading, multipath, and noise to a clean synthetic IQ signal to replicate the distortions encountered in real-world wireless propagation. Channel simulation bridges the gap between pristine synthetic data and the harsh reality of over-the-air collection.

• Multipath Rayleigh Fading: Simulates urban environments with non-line-of-sight reflections
• Rician Fading: Models scenarios with a dominant line-of-sight path
• Doppler Shift: Introduces frequency spreading for mobile transmitter/receiver scenarios
• Phase Noise: Emulates local oscillator imperfections

Modern channel simulators use tapped delay line models and Clarke's model to generate statistically accurate fading coefficients.

A data regularization technique that applies realistic channel impairments—such as phase rotation, noise addition, and fading—to synthetic or collected IQ samples to expand training dataset diversity. I/Q augmentation is the primary defense against overfitting in modulation classifiers.

• Phase Rotation: Uniform angular shift teaching rotational invariance
• Gaussian Noise Addition: Varies the effective SNR across training examples
• Time Stretching: Simulates sample rate mismatches
• Frequency Offset Injection: Introduces controlled CFO to improve robustness

Augmentation transforms a single synthetic signal into hundreds of realistic variants, dramatically reducing the need for expensive over-the-air data collection campaigns.

A fundamental channel model that adds a random noise signal with a flat spectral density and Gaussian amplitude distribution to the IQ stream, simulating thermal noise in the receiver. AWGN is the baseline impairment applied to all synthetic IQ generation pipelines.

• Noise Floor: Defined by kTB where k is Boltzmann's constant, T is temperature in Kelvin, and B is bandwidth
• SNR Range: Synthetic datasets typically span -20 dB to +30 dB SNR
• Es/N0 vs Eb/N0: Energy per symbol vs energy per bit normalization

While AWGN alone is insufficient for realistic training, it serves as the foundation upon which more complex impairments like multipath fading and adjacent channel interference are layered.

The sequence of signal conditioning steps applied to raw IQ samples—such as normalization, centering, and filtering—to create a standardized input tensor for a machine learning classifier. I/Q preprocessing ensures synthetic training data and real-world inference data share identical statistical properties.

• Z-Score Normalization: Transforms IQ samples to zero mean and unit variance
• I/Q Centering: Removes residual carrier frequency offset
• DC Offset Removal: Subtracts the mean to eliminate hardware bias
• Peak-to-Average Power Ratio (PAPR) Clipping: Prevents numerical overflow

Consistent preprocessing between synthetic training and production inference is critical—a mismatch in normalization statistics can silently degrade classifier accuracy by 10-20%.

A neural network design that processes IQ data natively as complex numbers, using complex-valued weights and activation functions to preserve the phase relationships inherent in the signal. Complex-valued networks are particularly well-suited for synthetic I/Q because they exploit the mathematical structure that real-valued models must learn implicitly.

• Complex Convolution: Applies complex filters where both real and imaginary parts are learned
• Complex ReLU: Activation function operating on magnitude and phase separately
• Wirtinger Calculus: Enables backpropagation through complex-valued operations

When trained on synthetic I/Q with precise phase labels, complex-valued networks can achieve superior modulation recognition accuracy at low SNR compared to dual-channel real-valued architectures.